/Anchor-based-hierarchical-reinforcement-learning

Anchor: The achieved goal to replace the subgoal for hierarchical reinforcement learning

Primary LanguagePython

This is a code implementation of the algorithm Anchor-based HRL (AHRL) presented in the following paper.

Anchor: The achieved goal to replace the subgoal for hierarchical reinforcement learning (Ruijia Li, Zhiling Cai, Tianyi Huang, William Zhu) https://www.sciencedirect.com/science/article/pii/S0950705121003919

AHRL is tested by MuJoCo and OpenAI gym. Networks are trained using PyTorch.

Usage

To train a policy on the Point Maze task by running: python main.py --env PointMaze

To see the performance of a policy trained on the Point Maze task by running: python test.py --env PointMaze

Acknowledgments

We would like to thank:

  • TD3. Our codebase is based on theirs.
  • DSC. Our environment Point Maze is based on their code.
  • HIRO. Our environment Ant Push is based on their code.
  • RLLAB. Our environment Double Inverted Pendulumis is based on their code.